vn-ocr-documents-eval v0.3

107 single-page Vietnamese documents for evaluating PDF / image → DOCX OCR pipelines. Six configs covering the full register matrix (formal + business + conversational + literary) plus real PD scans and synthetic receipts.

Config	n	Source	License
real	9	chinhphu.vn + hanoi.gov.vn signed scans	Public Domain (Luật SHTT VN, Điều 15)
formal	24	UDHR-vie articles + scan artifacts	CC0 (rendered) — UDHR text is PD
news_business	24	wiki_vi article openings + scan artifacts	CC-BY-SA 4.0 (Wikipedia VN)
conversational	24	tatoeba_vi sentence groups + scan artifacts	CC-BY 2.0 FR (Tatoeba)
literary	23	wikisource Truyện Kiều excerpts + scan artifacts	Public Domain
receipt	21	7 templates x 3 seeds (tax invoice, payment, expense voucher, transport ticket, donation, utility, medical)	CC0 1.0
contract	20	5 templates x 4 seeds (labor, rental, economic, service, loan)	CC0 1.0
form	20	5 templates x 4 seeds (resignation, residence cert, enrollment, leave, business registration)	CC0 1.0

What's new vs v0.1

• v0.1 was 100 % PIL-rendered synthetic content. That made convert_to_docx look like a 0.30 % CER tool — accurate for the synthetic, misleading for real scans.
• v0.2 ships 9 real public-domain Vietnamese government scans across central + provincial domains and 7 document types (Quyết định, Công văn, Nghị quyết, Thông tư, Thông báo, Kế hoạch + the synthetic receipts).
• The honest baseline number on real scans is 12.62 % whitespace-normalized CER — 42x worse than the v0.1 synthetic baseline. That's the gap a "real document OCR" pipeline actually needs to close.

Source diversity

Issuer	Source domain	Doc types
Thủ tướng Chính phủ	chinhphu.vn	Quyết định x 2
Văn phòng Chính phủ	chinhphu.vn	Công văn
Chính phủ	chinhphu.vn	Nghị quyết
Bộ Công Thương	chinhphu.vn	Thông tư
Bộ Công an	chinhphu.vn	Thông tư
UBND TP Hà Nội	hanoi.gov.vn	Quyết định, Thông báo, Kế hoạch
(synthetic)	—	Hoá đơn, Biên lai, Phiếu chi

Usage

from datasets import load_dataset

# Real PD scans (each row = one page + ground-truth text)
real = load_dataset("nrl-ai/vn-ocr-documents-eval", "real", split="test")
print(real[0]["text"][:120])
real[0]["image"].show()

# Synthetic receipts with scan artifacts
synth = load_dataset("nrl-ai/vn-ocr-documents-eval", "synthetic_scan", split="test")

For end-to-end PDF → DOCX evaluation, use the per-doc PDFs in docs/<doc_id>.pdf plus the flat metadata.jsonl (text field) which covers both configs.

Reference baseline — nom.convert.convert_to_docx

Latest in-house bench (source), Tesseract 5 (vie+eng pack) via the pdf_to_docx OCR-fallback path:

Config	CER (whitespace-normalized)	n
real	12.62 %	9
synthetic_scan	0.43 %	3
OVERALL	9.57 %	12

Throughput: ~0.7 docs/sec on a single CPU.

CER computed on whitespace-normalized strings (NFC, runs of whitespace collapsed to single space). The 12.62 % real-config CER reflects the combined difficulty of: (a) skewed scans with stamps, (b) admin abbreviations Tesseract's vie pack doesn't handle gracefully (signature blocks, "KT.", etc.), (c) low-contrast watermarks behind body text.

Per-doc CER ranges 4-22 % — real_qd_707_ttg is the cleanest scan (4 %), real_hanoi_tb_453_flag is the worst (22 %, short bold title with stamps).

Run yourself:

git clone https://github.com/nrl-ai/nom-vn
cd nom-vn
pip install -e ".[doc]"
python benchmarks/data/vn_documents_ocr_v2/_generate.py
python benchmarks/accuracy/bench_convert_documents_v2.py

Honesty notes

• n=12 is small. Use for smoke + regression checks; for adoption claims expand to 50-100 docs covering more issuers and date ranges.
• Ground truth is human-verified visual transcription — each text field was produced by directly reading the rendered page (not by trusting Tesseract output). Errors of judgment may remain in long numbered lists.
• Page 1 only. Some sources are multi-page; only page 1 is in the corpus. The sources/<id>_full.pdf files preserve the originals for future expansion.
• No PII. The chinhphu.vn / hanoi.gov.vn documents name public officials in their official capacity (PM, Ministers, Phó Chủ tịch UBND) — public record. Synthetic receipts contain only fictional names.

License

• The 9 documents in real config: Public Domain under Luật Sở hữu trí tuệ Việt Nam, Điều 15 (Vietnamese government works are not subject to copyright). Source URLs in metadata.
• The 3 documents in synthetic_scan config: CC0 1.0 Universal (synthetic content, no rights reserved).
• This README + the _generate.py script: CC0 1.0.

Citation

@dataset{nguyen_vn_ocr_documents_eval_v2_2026,
  author = {Nguyen, Viet-Anh and {Neural Research Lab}},
  title  = {{vn-ocr-documents-eval v0.2: Realistic Vietnamese
             scanned-document evaluation set with central + provincial
             government documents}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/nrl-ai/vn-ocr-documents-eval}
}

Maintained as part of the nom-vn project by Viet-Anh Nguyen (vietanh@nrl.ai) and Neural Research Lab.